F-Secure has identified 26 ways in which AI may harm consumers, as well as how to prevent these risks.
The new AI damage classification method tracks how damage spreads within the AI system, from misleading chatbot responses to compromised autonomous proxies, and identifies at which points defense measures can be implemented.
Los Angeles, November 10 (Reuters) / PRNewswire / -- Artificial intelligence is now influencing the way we search, communicate, make decisions, and manage our digital lives. But what kind of harm can it cause to people when it goes wrong? F-Secure The newly released AI Harms Taxonomy maps out 26 different types of consumer harm, tracking how each type of harm affects users. As a result, it clearly shows where protective measures can be implemented and who is best equipped to provide assistance.
This classification system was developed by Dr. Megan Squire, the Chief Threat Intelligence Researcher at F-Secure. It aims to address a blind spot in the risk discussions of AI. Most frameworks only list potential error-prone areas but do not further link these vulnerabilities to the systems and companies that could prevent them. The research at F-Secure goes a step further by detailing each layer between the threats that drive AI and their impact on consumers, and based on this, it indicates where defenses should be deployed.
Dr. Squire stated: "As the capabilities of the AI systems continue to enhance, the consequences of their mistakes could far exceed simply providing a poor answer. Those who develop these systems have the responsibility to understand how they may fail, who might be harmed in such cases, and to put in place protective measures before any damage occurs."
This classification reflects that the risks themselves are changing. In the past, criminals might have targeted individuals directly, but now AI is involved in the middle, which has introduced new participants and new ways of causing problems. An assistant acting on behalf of others could be manipulated to cause damage, and chatbots may also confidently mislead users. At a more fundamental level, the issues often stem from how AI processes instructions and memory, or which other services it connects to.
This study currently categorizes 26 types of damages into five relationship types and 19 intermediary layers, and has designed 24 defensive measures to address risks at different stages. Examples include voice and video impersonation generated by AI, agents making harmful financial decisions, leaking personal information, or taking actions in outdated contexts. The study also examines a type of attack where malicious instructions are hidden within content processed by the AI system, which may then affect the behavior of the agents or what chatbots tell users.
One of the key findings of this framework is that the responsibility for preventing AI from causing harm is distributed among multiple parties. Some risks can be handled by security software on individual devices; others must be addressed at a higher level, with companies that develop AI products and their underlying systems being responsible, including how AI stores information and which other services it connects to. Clarifying who is responsible for each step helps to determine where defenses should be placed and who must build them. This also undermines the notion that people can remain safe on their own simply by identifying suspicious content or repeatedly verifying the information provided by AI.
As AI shifts from "generating information" to "taking action based on information," the question of who is responsible becomes more urgent. If a chatbot gives an incorrect answer, it may mislead users, but at least the users are still involved in that interaction; however, an agent that has access to users' accounts and their associated services could cause real consequences without such back-and-forth confirmation. Protective measures need to be built around what these systems actually do, and this begins with understanding the differences between chatbots and agents.
Build protection for AI proxy
This research is part of a broader effort by F-Secure to build trust for the era of AI. F- Secure Trust announced this alongside its strategic alliance with Quantum AI Lab QuTwo, as part of the company's commitment to protecting users in a world where AI agents are increasingly acting on behalf of people. TrustGuard is aimed at helping to safeguard people's security during interactions in AI, while TrustPath focuses on protecting their digital journey from discovery to delivery. AI Harms Taxonomy provides a research foundation for these efforts, identifying the risks that protective measures need to address, and also makes it clear that effective protection also depends on the companies that build and operate AI systems.
Research for the era of AI
Today's digital threats are no longer just about fraud. The team at Illuminate Research studies how people and technology interact in reality, exploring the complex realities of trust, deception, and online resilience in the AI era. AI Harms Taxonomy can be obtained at Illuminate Research Hub.
About F-Secure
F-Secure is a people-oriented company dedicated to enhancing consumer online security experiences, driven by AI. With 38 years of experience in addressing digital threats, the company helps digital service providers transform trust into a high-value growth engine. In a world plagued by fraud driven by AI, it protects their customers while enabling them to lead better digital lives. The company safeguards billions of digital interactions each year, protecting tens of millions of consumers worldwide, and creates over $10 billion in value for its partners, achieving a proven large-scale impact.
News contact person:
- Meghan Sawyer, Senior Public Relations Manager, [ email protected ]
- Joel Latto, Threat Consultant and EMEA Public Relations Manager, [ email protected ]
The above information is provided by Cision: http :// news.cision.com
